The mdx system consists of generic CPU nodes, GPU acceleration nodes, and storages. Resources are provided through virtual machines.
Compute Nodes
| Generic CPU Nodes | GPU Acceleration Nodes | |
|---|---|---|
| Model | PRIMERGY CX2550 M6 | PRIMERGY GX2570 M6 |
| CPU (per Node) | Intel Xeon Platinum 8368 Processor (38 core, 2.4 GHz) × 2 | Intel Xeon Platinum 8368 Processor (38 core, 2.4 GHz) × 2 |
| Memory | 256 GiB | 512 GiB |
| GPU (per Node) | NVIDIA Tesla A100 × 8 (GPU Memory 40 GiB) | |
| Number of Nodes | 368 | 40 |
| Total Theoretical Computational Performance (Double Precision) | 2.1 PFLOPS | 6.4 PFLOPS |
| Total Theoretical Computational Performance (Single Precision) | 6.7 PFLOPS | |
| Total Theoretical Computational Performance (Half Precision) | 100.7 PFLOPS |
Storage
| Model | Main Applications | |
|---|---|---|
| Virtual Disk | IntelliFlash HD2160 | Storage area for virtual machine OS |
| High-Speed Storage | DDN ES400NVX (SSD-based, 252 GByte/sec) | /fast. Shared directory between virtual machines within a project. For data requiring fast I/O, such as intermediate data during computation. |
| Large-Capacity Storage | DDN ES7990X (SSD + HDD Hybrid, 157.5 GByte/sec) | /large. Shared directory between virtual machines within a project. For data requiring large storage capacity, such as large-scale training data. |
| Object Storage | DDN ES7990X (SSD + HDD Hybrid, 63.0 GByte/sec) | S3-compatible shared storage. For data sharing outside of mdx. |
Virtual Machines
In mdx, you can create custom virtual machines to fit your specific needs. Virtual machines are built in units of CPU and GPU packs.
| CPU Pack | GPU Pack | |
|---|---|---|
| CPU Cores | 1 Virtual Core | 18 Virtual Cores |
| GPUs | 1 GPU | |
| Memory | 1.51GB | 57.60GB |
| GPU Memory | 40GB | |
| Maximum # Packs Assignable to 1 Virtual Machine | 152 Packs | 8 Packs |
| Total Theoretical Computational Performance (Double Precision) | Approx. 38.35 GFLOPS | Approx. 20.2 TFLOPS |
| Total Theoretical Computational Performance (Single Precision) | Approx. 20.9 TFLOPS | |
| Total Theoretical Computational Performance (Half Precision) | Approx. 315 TFLOPS |